PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine Learning for Scent: Learning Generalizable Perceptual Representations of Small Molecules

Benjamin Sanchez-Lengeling, Jennifer N. Wei, Brian K. Lee, Richard C. Gerkin, Alán Aspuru-Guzik, Alexander B. Wiltschko

arXiv:1910.10685v2stat.MLcs.LGphysics.chem-ph

TL;DR

Predicting odor from molecular structure remains a difficult, decades-old QSOR problem. The authors train graph neural networks on an expert-labeled molecule dataset and evaluate the learned embeddings. They report odor-space structure and strong results on related prediction tasks.

  • Problem

    QSOR—predicting a molecule’s odor from its structure—remains difficult and comparatively underexplored despite decades of work.

    Source: §1 · p. 1

  • Method

    The authors curate expert-labeled odor descriptors and train graph neural networks to predict them from molecular graph structure.

    Source: §1 · p. 2

  • Results

    The model outperforms prior QSOR methods, learns embeddings organized by perceptual similarity, and performs strongly on two related transfer tasks.

    Source: §1 · p. 2, §6.2 · p. 7, §6.3 · p. 8

  • Takeaways & Limitations

    The authors suggest the learned embeddings may be useful for rational design of new odorants.

    Source: §7 · p. 9

Abstract

from arXiv · show

Predicting the relationship between a molecule's structure and its odor remains a difficult, decades-old task. This problem, termed quantitative structure-odor relationship (QSOR) modeling, is an important challenge in chemistry, impacting human nutrition, manufacture of synthetic fragrance, the environment, and sensory neuroscience. We propose the use of graph neural networks for QSOR, and show they significantly out-perform prior methods on a novel data set labeled by olfactory experts. Additional analysis shows that the learned embeddings from graph neural networks capture a meaningful odor space representation of the underlying relationship between structure and odor, as demonstrated by strong performance on two challenging transfer learning tasks. Machine learning has already had a large impact on the senses of sight and sound. Based on these early results with graph neural networks for molecular properties, we hope machine learning can eventually do for olfaction what it has already done for vision and hearing.

1 Introduction

QSOR—predicting odor from molecular structure—remains a difficult, comparatively underexplored problem. This work applies graph neural networks to expert-labeled odor data and reports embeddings organized by perceptual similarity and useful for related prediction tasks.

1 Introduction

  • QSOR modeling has received comparatively little attention during growth in machine learning for molecular properties, despite being a 70+ year-old problem spanning several fields.The paper distinguishes QSOR from quantitative structure-activity relationships used in medicinal chemistry.Source: §1 · p. 1
  • Odor perception involves 300–400 olfactory receptor types, and improved olfaction models may aid synthetic-odorant discovery and contribute to understanding sensory perception.The introduction also connects synthetic-odorant discovery with reducing the ecological impact of harvesting natural products.Source: §1 · p. 1
  • Figure 1 illustrates why structure alone is an imperfect guide to odor: small structural changes can render a molecule odorless, while large changes can leave its odor largely unchanged.The examples are attributed to Ohloff, Pickenhagen and Kraft [28].Source: Figure 1 · p. 2
  • The authors train graph neural networks on expert-labeled odor descriptors and report embeddings that cluster molecules by perceptual similarity globally and locally, with utility on related tasks.They describe the learned representation as a general-purpose representation of the structure–odor relationship, anticipated to aid rational molecular design and screening.Source: §1 · p. 2

2 Prior Work in QSOR: A Decades-Long Pursuit

Prior QSOR work spans decades, but earlier studies often used few odor descriptors and molecules. Traditional approaches represent molecules with hand-crafted features or fingerprints and use conventional prediction models.

2 Prior Work in QSOR: A Decades-Long Pursuit

  • Earlier QSOR studies typically used fewer than ten odor descriptors and tens, rarely hundreds, of molecules; activity cliffs make structure-to-odor prediction difficult.Lyral illustrates that structural neighbors need not be perceptual neighbors, and perceptual neighbors may have little structural similarity.Source: §2 · p. 2
  • The DREAM challenge evaluated 49 untrained panelists’ ratings of 476 molecules on 21 odor attributes, with winning models relying mainly on Dragon features or Morgan fingerprints and random forests.The paper uses these methods as baselines.Source: §2 · p. 2

2.1 Classic Approaches to Featurizing Molecules and Modeling Their Properties

  • Classic approaches convert molecules into fixed-length vectors of hand-crafted features and feed them to models such as random forests or fully connected neural networks.Dragon and Mordred aggregate thousands of molecular features; Mordred was selected as open source, with no appreciable performance difference from Dragon reported.Source: §2.1 · pp. 2–3, §2.1.1 · p. 3

2.1.2 Molecular Fingerprints

  • Molecular fingerprints encode graph environments into fixed-length vectors; this work compares bit-based path-descriptor fingerprints and count-based Morgan fingerprints using random forests.Fingerprint settings include maximum topological radius and vector size, which affect represented fragment size and hash-collision likelihood.Source: §2.1.2 · p. 3

3 Graph Neural Networks

Graph neural networks let models operate directly on molecular graphs rather than requiring regularly shaped inputs. Here, the model predicts odor descriptors from graph features and saves an intermediate representation as an odor embedding.

3 Graph Neural Networks

  • Graph neural networks enable machine learning on irregular inputs such as graphs, with atoms treated as nodes and bonds as edges in molecular applications.The paper describes GNNs as learnable, permutation-invariant transformations that produce fixed-length vectors for further processing.Source: §3 · p. 3, §3.1 · p. 3

3.1 Graph Neural Networks for Predicting Molecular Properties

  • The molecular GNN transforms graph features through message-passing layers, reduces the final layer to a vector, and uses fully connected layers to predict odor descriptors.The model schematic shows atoms, bonds, and connectivities as input features and a fully connected network as the descriptor-prediction stage.Source: §3.1 · pp. 3–4, Figure 2 · p. 4

3.2 Learned Graph Neural Network Embeddings

  • The authors save activations from the penultimate fully connected layer as a fixed-dimension odor embedding for testing its organization and downstream utility.They treat the embedding as a learned molecular representation whose usefulness is evaluated in later sections.Source: §3.2 · p. 4

4 A Curated QSOR Dataset

The authors assemble an expert-labeled, multi-label dataset of single-molecule odorants from two databases. Its descriptors co-occur in meaningful groups, motivating joint prediction across all 138 retained descriptors.

4 A Curated QSOR Dataset

  • The curated dataset contains 5,030 molecules from GoodScents and Leffingwell, with 2,317 overlaps, and retains 138 descriptors after filtering for at least 30 representative molecules.The sources contribute 3,786 and 3,561 molecules, respectively; overlapping molecules inherit the union of their descriptors.Source: §4 · p. 4
  • The dataset is multi-label and perfumery-biased, with uneven descriptor frequencies and label specificity; Figure 3 summarizes descriptor frequency, label density, and descriptor co-occurrence.The authors note that common labels such as fruity are more frequent than specific labels such as pineapple, and the dataset is biased away from malodorous compounds.Source: §4 · p. 4, Figure 3 · p. 5
  • Odor descriptors show strong co-occurrence patterns, including dairy, fruity, and bakery groups, so the authors train one GNN to predict all 138 tasks at once.This joint prediction contrasts with prior approaches that often trained one model per descriptor.Source: §4 · p. 5

5 QSOR Prediction Performance Benchmark

The benchmark treats odor-descriptor prediction as multi-label classification and compares GNNs with random forests and k-nearest neighbors across several molecular featurizations.

5 QSOR Prediction Performance Benchmark

  • The benchmark compares GNN, random forest, and k-nearest-neighbor models using bit-based RDKit fingerprints, count-based Morgan fingerprints, and Mordred features.Non-graph fully connected neural networks on bit- and count-based fingerprints performed indistinguishably from random forests.Source: §5 · p. 5
  • Several metrics are reported because each can highlight different performance characteristics.Source: §5 · p. 5
  • Mean AUROC averaged across odor descriptors is the primary comparison for the remaining analysis; descriptor-level AUROC is shown in Figure 4.Source: §5 · p. 5

6 Evaluating Odor Embeddings

The section evaluates whether GNN embeddings organize molecules by odor at global and local scales, then tests whether that organization transfers to new descriptors and another olfaction task. The reported results show perceptual structure and promising transfer, while DREAM performance is statistically indistinguishable from the challenge winner.

6 Evaluating Odor Embeddings

  • The evaluation asks whether perceptually similar odors are nearby globally and whether molecules with similar odor percepts cluster locally.The embeddings are taken from a trained GNN’s penultimate layer; the best model maps each molecule to a 63-dimensional vector, visualized using PCA and KDE.Source: §6 · p. 5, §6.1 · pp. 5–6

6.1 Examining the Global Structure of a Learned Odor Space

  • GNN embeddings organize odor descriptors globally: specific labels cluster in regions, related labels form hierarchies, and GNN outperforms RF-cFP on nearly all descriptors.The embedding reflects the block structure of odor relationships, unlike molecular fingerprints; the visualization also shows broad labels subsuming more specific ones.Source: §6.1 · pp. 6–7, Figure 4 · p. 6, Figure 5 · p. 7

6.2 Evaluating the Local Structure of a Learned Odor Space

  • AUROC = 0.818 (95% CI [0.806, 0.830]) for GNN embeddings versus AUROC = 0.782 (95% CI [0.773, 0.797]) for bit-based fingerprints on k-nearest-neighbor prediction.Nearest-neighbor retrieval with GNN embeddings also yields molecules described as more perceptually similar to the source molecule.Source: §6.2 · p. 7, Figure 6 · p. 8

6.3 Transfer Learning to Previously-Unseen Odor Descriptors

  • GNN embeddings significantly outperform Morgan fingerprints and Mordred features on held-out odor prediction, though they perform slightly worse than a GNN trained on the target odor.The test ablates one odor descriptor and trains a random forest on embeddings learned from the remaining N −1 descriptors.Source: §6.3 · p. 8, Figure 7 · p. 8

6.4 Generalizing to Other Olfaction Tasks: the DREAM Olfaction Prediction Challenge

  • Mean Pearson’s r = 0.55 for a random forest on GNN embeddings versus 0.54 for the DREAM-winning model, but their r and R2 performance is statistically indistinguishable.Source: §6.4 · p. 9, Figure 8 · p. 9

7 Conclusion

The authors assembled a large expert-labeled dataset and trained a graph neural network for structure-to-odor prediction. They report state-of-the-art QSOR results, meaningful embedding structure, and downstream utility, suggesting possible use in rational odorant design.

7 Conclusion

  • The model achieved state-of-the-art results on QSOR against field-recognized baselines.The authors also report that its embeddings capture meaningful structure at both local and global scales.Source: §7 · p. 9
  • The authors assembled a novel, large dataset of expertly labeled single-molecule odorants and trained a graph neural network to predict smell from molecular structure.Source: §7 · p. 9
  • The learned embeddings were useful in downstream tasks, a property the authors describe as rare among modern machine-learning models and chemistry data.Source: §7 · p. 9
  • The authors suggest the model and embeddings might be useful in the rational design of new odorants.Source: §7 · p. 9

Supporting Information

The supporting information reports model-selection and evaluation procedures, then provides supplementary comparisons of descriptor performance and embedding structure. It also describes transfer results against the DREAM challenge model.

Hyperparameter Tunning and GNN Architecture

  • The GNN hyperparameters were selected by five-fold cross-validation and 500 random-search trials, while RF and KNN baselines used tuned feature and model configurations.The KNN configuration used a 4096-bit RDKit fingerprint with radius 6 and k = 20, with distance-weighted predictions.Source: Hyperparameter Tunning and GNN… · p. 12
  • MPNNs and GCNs performed similarly and both significantly outperformed all baseline models; the authors report GCN analyses because that architecture is simpler.Source: Hyperparameter Tunning and GNN… · p. 12
  • Second-order iterative stratification was used to preserve ratios of label pairs across train, test, and validation splits for the highly unbalanced multi-label task.Source: Hyperparameter Tunning and GNN… · p. 12

Confidence Intervals

  • The 95% AUROC confidence intervals were computed from 1,000 bootstrap resamples of the test set, with the training set and model held fixed.Source: Confidence Intervals · p. 12

Table of Per-Descriptor Results

  • The per-descriptor table reports AUROC and AUPRC for the GNN and random forest with count-based Morgan fingerprints.Source: Table of Per-Descriptor Results · p. 13
  • The GCN embedding’s correlation with odor-label co-occurrence was 0.43, compared with 0.22 for Morgan fingerprints.The matrices use log-scaled colors and are normalized so each row and column sums to 1.Source: Figure S1 · p. 16
  • The GNN and fingerprint embedding distances were evaluated against Jaccard label distance using Euclidean, cosine, and Jaccard metrics across the specified spaces.Source: Figure S2 · p. 16
  • The fingerprint embedding did not reflect odor-descriptor hierarchy: low-co-occurrence labels overlapped, unlike the hierarchy reported for GNN embeddings.The figure shows PCA-reduced fingerprints and KDE contours for positive labels.Source: Figure S3 · p. 17

Related papers

Loading 1910.10685v2…